U-Net-Based Classification of Patient Sleep Postures Using IMU-Derived RGB Representations

Abstract

Patients confined to bed for extended periods must frequently change sleeping posture to prevent pressure ulcers, which are difficult and undesirable to treat. In this study, three IMU sensors were placed on 108 bedridden patients to collect data for five different postures, resulting in 1,800,000 data points per sensor. These were converted into Eulerx, Eulery, and Eulerz values. The goal was to detect the sleep posture using a single IMU sensor. Four cases were defined: Case 1 used only IMU1, Case 2 only IMU2, Case 3 only IMU3, and Case 4 combined all three. Euler signals were converted into RGB images, framing the problem as an image classification task. A total of 2000 images were used, with 400 for training and 100 for testing in each case. A U-Net model was applied, achieving high IoU scores: 98.00%, 99.56%, 99.72%, and 98.20% respectively. Accuracy scores were 98.98%, 99.78%, 99.86%, and 99.09%, confirming U-Net's effectiveness.